基于均值方差预测与掩码全变分损失的高光谱图像超分辨率

Hyperspectral Image Super-Resolution Based on Mean Variance Prediction and Masked total Variation Loss

摘要: 高光谱图像因其丰富的空间与光谱信息，在农业、遥感、环境监测等领域具有重要应用价值。然而，受成像硬件限制，其空间分辨率往往较低。因此，众多深度学习方法被提出用于高光谱图像超分辨率。然而，大多数深度学习方法是从高光谱图像中切出部分进行训练，在剩余部分图像上进行仿真，这在现实中是不适用的，但是神经网络单波段地重构图像又会导致光谱偏差。其次，在深度学习方法的损失函数中，全变分正则化作为一种常用的损失函数，旨在约束图像以使图像平滑，但该约束也会导致图像中的纹理变得模糊，导致超分辨率结果不佳。本文提出一种基于均值方差预测与波段生成的网络架构以及使用掩码全变分损失的高光谱图像超分辨率方法，通过灰度图像进行训练，在提升空间分辨率的同时有效保持光谱特征。实验表明，所提方法在PSNR、SAM指标上优于使用传统的深度学习超分辨率方法，并在光谱保真度与空间细节恢复上表现出显著优势。

Abstract: Due to its rich spatial and spectral information, hyperspectral images have important application value in agriculture, remote sensing, environmental monitoring and other fields. However, limited by imaging hardware, its spatial resolution is often low. Therefore, many deep learning methods have been proposed for hyperspectral image super-resolution. However, most deep learning methods cut out part of the hyperspectral image for training and simulate on the remaining part of the image, which is not applicable in reality, but making the neural network reconstruct the image in a single band will lead to spectral deviation. Secondly, among the loss functions of deep learning methods, total variation regularization is a commonly used loss function that aims to constrain the image to smooth the image, but this constraint also causes the texture in the image to become blurred, resulting in poor super-resolution results. In this paper, we propose a hyperspectral image super-resolution method based on mean-variance prediction and band generation network architecture and masked total variation loss. By training on grayscale images, the spatial resolution is improved while the spectral features are effectively preserved. Experiments show that the proposed method is superior to the traditional deep learning super-resolution method in terms of PSNR, SAM indicators, and shows significant advantages in spectral fidelity and spatial detail recovery.